NEWS.md

In rtemis: Schema-Driven Machine Learning and Visualization

rtemis news

1.4.2

• The standalone Plotly, networkD3, threejs and leaflet draw_*() functions move to the 'rtemis.dplot' package on r-universe, with their themes (theme_*(), choose_theme(), available_themes()), palettes (get_palette()), color utilities (col2grayscale(), color_adjust(), previewcolor()), available_draw(), uniprot_get(), xtdescribe(), the xt_example data and the rtemis_theme, rtemis_palette and rtemis_font options. 'rtemis.draw' provides their replacements. rtemis keeps draw_linad(), which takes an 'rtemis.draw' Theme, and roc_curve().
• setup_PCoA() configures principal coordinates analysis (classical multidimensional scaling) on a choice of dissimilarities; it applies to new data through Gower's out-of-sample formula, so it can be train()'s decomposition step, and does not reconstruct the input.
• setup_MDS() configures metric or nonmetric multidimensional scaling through 'vegan' with multiple starts; it cannot be applied to new data and does not reconstruct the input.
• review() assesses a trained supervised model, single-split or resampled: it reports sample sizes, every training and test metric, and comparisons with a baseline predictor, with confidence intervals and tests for a single split and a description of the variation between resamples for a resampled model, then notes small samples, many predictors, and signs of overfitting.
• ai_review() asks a language model, through 'rtemis.llm', to write a summary, evaluation, next steps and caveats from a review(), statements citing the review's finding codes, checked against the review; the result keeps the review and records the model, settings, prompt and a hash of the review.
• A fitted model records the value its backend chose for a hyperparameter left unset, so the model and its record state every value the fit used: for example mtry for ranger, nk for MARS, the loss and optimizer settings for MLP, and the engine settings for LINAD.
• LightRuleFit stops with a corrective error when its boosting stage makes no split.
• setup_TabNet() rejects importance_sample_size when skip_importance is TRUE.
• setup_CART(cost =) and setup_Ranger(class_weights =) reach their backends; class weights named by class are matched to the outcome's levels.
• setup_Ranger() drops case_weights, which ranger did not receive; per-case weights are passed with train(weights =).
• setup_GLMNET() drops offset, which failed at prediction and under resampling.
• setup_LINAD() and setup_LINADForest() take root_forward_stop and root_node_test, the root model's stopping rule and slopes test; unset uses the node-level value.
• Variable importance (get_varimp()) holds one measure per name, each stating what it quantifies (split gain, permutation, coefficient, ...), the cases it was computed on, whether it is signed or depends on predictor units, how it ranks predictors, and a description for the fitted model; the measures replace the single table in @data, and varimp_table() combines them into one data.frame. plot_varimp() ranks bars by the measure's direction (a cross-validated risk smallest first). Linear SVM weights are oriented toward the positive class, and a forest fitted with importance = "none" has no importance.
• writeup() writes the Methods and Results sections for a trained supervised model from the model and its review(), every number traceable to the field it was read from, with references to R, rtemis, the fitting packages and the statistical methods, a main hyperparameter table (primary hyperparameters and every one tuned, specified or resolved during fitting; include_hyperparameters replaces the primary list) and a supplementary table of every hyperparameter with its value, how it was chosen and the values tried, and a list of what the model does not record.
• write_text() writes a writeup(), review() or ai_review() as Markdown.
• Stratified resampling (setup_KFold(), setup_StratSub(), setup_StratBoot()) stratifies a categorical variable by its levels; with more than four levels it grouped adjacent levels together.
• Metric labels print AUC, PPV and NPV in capitals.
• LightRF predictions are no longer pulled toward 0 when some trees cannot split, as happens on small samples: such trees now predict their sample's center. Training reports how many trees could not split.
• setup_SuperConfig() and setup_SuperConfigLive() reject a decomposition that cannot be applied to new data, as reading a supervised config already did.
• Shared results identify implementation-specific learner, execution, preprocessing, and resampler configs by schema and read their typed settings without inserting omitted defaults.
• write_config() and write_record() write numbers with 15 significant digits instead of rounding to four decimal places.
• calibrate() takes execution_config; calibration models train serially when it is unset.
• train() and the setup_*() functions spend less time reading property declarations and default policies; a single CART fit on iris takes about a third of the time.
• Each parallel task receives only its own resample or bag; LINADForest and bias_variance() pass training data to local workers through shared memory.
• GAM classification trains on a data.table, which failed for every outcome, and on a multiclass outcome, using mgcv's multinomial family. A multiclass GAM reports no variable importance or standard errors, and explain() refuses its GAMTerms estimator; explanation_methods() reports GAM's exactness as dependent on the fit.
• LinearSVM and RadialSVM probabilities are those of the positive class for a binary outcome and follow the outcome's level order for a multiclass one. Binary probabilities were those of the negative class when the first training case belonged to the positive class.
• setup_GridSearch() with maximize unset maximizes every metric for which a higher value is better (accuracy, balanced accuracy, sensitivity, specificity, PPV, NPV, F1, AUC, R², concordance) and minimizes the others; it minimized AUC, sensitivity, specificity, PPV, NPV and F1.
• se() is exported: standard errors of the fit for GLM, GAM, BART and LINADForest models.
• get_metrics() returns the metrics a supervised model, resampled model, clustering or decomposition stores for one sample; calibrated = TRUE selects a calibrated classifier's calibrated metrics.
• $ and [[ read the properties of resampled models, resampled metrics, tuners, LightRuleFit models and bias_variance() results, with tab completion, and [[ reads any property of a decomposition, as $ does. On these objects and on models and clusterings, a name that is not a property raises an error listing the properties.
• plot_varimp() draws a SHAP explanation's mean absolute contributions.
• plot() draws a fitted supervised model's default view: true against predicted values for regression, confusion matrices for classification, and metric distributions for resampled models. present() prints the model's description and returns the same view; plot_roc() draws ROC curves.
• Model plots color each data split the same way: training teal, validation purple, test magenta. Confusion matrices shade correct predictions blue and incorrect predictions orange.
• Imputation fitted on training data is applied to new data: apply_preprocessor(), and train() on validation, test and resampled test data, impute with what was learned from the training data rather than from the new data.
• Mean and mode imputation (impute_type = "meanMode") stores the value imputed into every feature in the new config properties impute_numeric_values and impute_level_values.
• Mean and mode imputation keeps each feature's type. impute_numeric ("mean" or "median") and impute_categorical ("mode") replace impute_discrete and impute_continuous. Integer features take the mean or median rounded half to even, logical features their more frequent value (FALSE on a tie), and factor and character features their mode, a tie going to the first level or the first value in code point order. Integer features took their mode, a tie going to the value appearing last. A supplied value that a feature's type cannot hold is an error.
• missRanger imputation keeps its fitted random forests in the Preprocessor (@imputer) and applies them with predict(); features with no missing training values take a mean or mode value on new data. All impute_missRanger_params reach missRanger, which received only pmm.k.
• micePMM imputation is available to preprocess() only: setup_SupervisedPreprocessor() does not offer it, and apply_preprocessor() refuses a preprocessor that uses it.
• setup_SupervisedPreprocessor() imputes with mean and mode values by default (impute_type = "meanMode"); missRanger, fitted on every training set of a train() run, is chosen explicitly. The Imputation sections of setup_SupervisedPreprocessor() and setup_Preprocessor() describe each method's cost and limits.

Supervised result validation

• Outcomes, predictions, and classification probabilities must have matching row counts within each sample.
• Categorical outcomes and predictions must preserve the training class levels and their order; probability matrices require one column for binary classification or one per class for multiclass classification.

1.4.1

• Model plotting and present() use rtemis.draw, including session timelines and SVG export; standalone Plotly draw_*() functions remain available.
• Require R >= 4.4.0 for the rtemis.draw backend.

• fit_predict() trains one regression model from an algorithm name and setup arguments (params) and predicts new data; rtemis.draw uses it for fit = "<algorithm>" and fit_params in scatter plots.

• TabNet and MLP train and predict with the algorithm workers resolved from execution_config (n_workers_algorithm) rather than every core.

• train() prints one resources line: the compute device in use (including an Apple silicon GPU), the execution backend, the worker ceiling, and each level's share.
• decomp(), cluster(), setup_DecomposeConfig() and setup_ClusterConfig() take execution_config: its seed seeds the fit, and a threaded algorithm (UMAP, tSNE) runs on its workers; decomposition_traits() gains threaded.
• setup_tSNE(num_threads =) is deprecated; tSNE takes its threads from the execution config.
• setup_Autoencoder() configures a torch autoencoder decomposition, denoising when input_noise or input_dropout is set; it applies to new data, so it can be train()'s decomposition step, and reconstruct() returns its reconstruction in the units of the data.
• setup_VariationalAutoencoder() configures a torch variational autoencoder decomposition; beta weighs the KL divergence (1 is the standard VAE, larger values a beta-VAE), and its components are the latent means.
• decomp_metrics() takes execution_config.
• NMF with nrun greater than 1 runs.
• Execution configs take device: "cpu", "cuda", "mps" (Apple silicon GPU) or "opencl", or setup_CUDA(ids =) to name GPUs. Unset uses CUDA where available and the CPU otherwise; the Apple silicon GPU runs only when named. An algorithm that cannot use the requested device runs on the CPU and the resources line says so.
• setup_LightGBM(device_type =), setup_LightRF(device_type =) and setup_TabNet(device =) are deprecated; set device in the execution config.
• predict(), apply_decomp() and reconstruct() take execution_config; predictions use its threads, or the host's default worker count, never the count a model was trained with.
• UMAP runs on the execution config's algorithm threads, which respect the two-core limit of a CRAN check, rather than half the hardware threads.
• setup_SpectralRBF(), setup_SpectralLaplace(), and setup_SpectralLocal() configure spectral clustering through 'kernlab', one per kernel; setup_Nystrom() enables the Nystrom approximation.
• setup_ASWCriterion(), setup_CHCriterion(), and setup_MultiASWCriterion() configure how setup_PAMK() selects the number of clusters.
• setup_DBSCAN(approx =) is a nonnegative number, the approximation factor of the neighbor search (0, the default, searches exactly); it was declared logical, and TRUE ran as an approximation factor of 1.
• setup_ClusterConfig() and setup_DecomposeConfig() no longer take algorithm; the clustering or decomposition config names it, and cluster() and decomp() reject an algorithm that disagrees with a supplied config.
• Clustering configs take features, the columns to cluster on.
• cluster() and decomp() use every numeric column when the config names no features, and record the columns used.
• Shared results identify implementation-specific configs by schema and read their typed settings without inserting omitted defaults.
• Implementation-specific schema paths include the language namespace; shared result paths remain unqualified.
• The defaults artifact format accepts producer-owned namespaces while retaining separate schema declarations and resolution policies.
• write_result() and read_result() support portable result JSON with optional Parquet payloads, lazy loading, and integrity checks.
• Resampled results retain successful fold identifiers and all requested training-row splits, including failed folds and bootstrap repetitions.
• Supervised, Regression, Classification, and their resampled counterparts publish shared result schemas with typed outcomes, predictions, metrics, and runtime-only fitted state.
• Categorical result values preserve class order and missing positions through explicit levels and codes.
• Matrix property declarations support numeric cell bounds and cell-level nullability.
• Typed array decoding preserves missing-cell positions and rejects invalid cell types and ragged matrices.
• Schema publication metadata identifies the generating language, contract scope, registry domain, and standalone parent class.
• Implementation and Provenance record package and language identities independently of schema authorship.
• VariableImportance publishes a shared report schema with named numeric measures and null values for unavailable results.
• JSONSchema_to_S7() reconstructs standalone inheritance and runtime-only properties from schema artifacts, and rejects a parent other than the one the schema declares.
• Reading a table or object with a field its schema does not declare fails with an error naming the field.

1.4.0

• setup_SerialExecution(), setup_FutureExecution(), and setup_MiraiExecution() configure their respective execution backends, replacing setup_ExecutionConfig().
• JSONSchema_to_S7() reconstructs declaration defaults, typed references, and input policies from a supplied artifact graph.
• Integer property declarations support exclusive lower and upper bounds.
• ClusteringMetrics records clustering measures with a status explaining each missing value.
• HardClustering and SoftClustering distinguish hard assignments from results with membership probabilities.
• setup_GMM() configures Gaussian mixture models through 'mclust', with optional selection of the number of components.
• setup_PAM() and setup_PAMK() configure medoid clustering through 'cluster' and 'fpc'.
• setup_HOPACH() configures hierarchical clustering through 'hopach'.

1.3.9

• Config families publish flat documents with a discriminator and the selected variant's settings.
• A separate authoring artifact identifies properties reserved for the host.
• partition() creates auditable train/test splits with random, time, group, or predefined partition configurations.
• setup_KFold(), setup_StratSub(), setup_StratBoot(), setup_Bootstrap(), setup_LOOCV(), and setup_Custom() configure individual resampling methods. setup_Resampler() remains available with a deprecation warning naming its replacement.
• SuperConfigPaths and SuperConfigTabular share common supervised configuration properties.
• is_wire_hyperparameters_set() identifies a serialized set of named learner configurations.
• Supervised records reference session graphs stored as Parquet sidecars.
• DataRef records a sidecar's path, encoding, digest, size, and optional table dimensions.

1.3.8

• ingest() reads supported data formats using format-specific configurations and writes Parquet output with a record of the reading decisions.
• Ingestion configurations accept explicit column type declarations.
• A published profile fixture pairs Parquet datasets with their expected data profiles.
• Supervised configurations expose character-to-factor conversion for file input.
• setup_SupervisedPreprocessor() configures preprocessing that preserves cases and feature identity during supervised resampling.
• Published checks, algorithm traits, and a shared conformance corpus support validation across implementations.
• train(preflight = TRUE) checks the configuration against the data before training.
• data_profile() records dataset dimensions, column types, missingness, distinct values, and level counts.
• validate_config() returns structured findings for configuration and data checks, including corrective patches where available.
• Diagnostic and Diagnostics have published schemas.
• check_data() reports distinct value counts per column.

1.3.7

• setup_LINAD() configures the Linear Additive Tree, a decision tree with a linear model at every node whose leaf coefficients are the sum along the path.
• LINAD recovers a pure linear model at max_leaves = 1 and a CART-equivalent tree with constant nodes at gamma = 0.
• setup_LINAD(split_search = "exhaustive") scores each split by the loss after fitting both child models.
• LINAD selects its number of leaves on dat_validation; force_max_leaves = TRUE keeps the full tree, and patience stops growth early.
• setup_LINAD() takes gamma, root_learning_rate, constant_rule, line_search, node_model {"forward", "ridge", "elasticnet"}, node_test, split_criterion, split_binning, and split_bin_type.
• setup_LINAD(split_features = , linear_features = , global_features = ) restrict which features may split, which enter the node models, and which have one coefficient shared by every leaf.
• get_varimp() on LINAD reports importance, the linear effect, and split_gain, the partitioning effect.
• draw_linad() draws a fitted LINAD tree as an interactive hierarchy with per-node coefficient tables; show_changes marks global, changed, and inherited coefficients.
• setup_LINADForest() configures a bagged ensemble of Linear Additive Trees, each sized on its own out-of-bag cases, with mtry_tree and mtry_split feature sampling.
• se() on a LINADForest returns infinitesimal jackknife standard errors.
• A fitted LINADForest reports out-of-bag predictions and metrics.
• LINADForest parallelizes across trees with results independent of worker count and backend.
• setup_GLMTree() configures model-based recursive partitioning through 'partykit', with separate regressors and partitioning_variables.
• plot_learning() and get_learning_curve() show loss against training progress for algorithms that train in steps.
• train() accepts a named list of hyperparameter configurations for one algorithm and tunes across all of them; mod@hyperparameters@variant names the selected configuration.
• tuning_grid() on a list of configurations returns the union of their grids with a .variant column.
• Configs serialize a set of hyperparameter configurations as {"variants": {...}}.
• bias_variance() decomposes a learner's error into bias and variance over resamples at a fixed test set; true_values measures bias against a known true function.
• Case weights must have one finite, non-negative value per case and a positive mean.
• Volcano and Manhattan plots share one package-wide pair of colors for positive and negative effects.

1.3.6

• Configs store the paths they are given, without resolving them against the working directory or expanding ~.
• predict() on a multiclass resampled classification returns an n x k probability matrix for type = "avg" and a list of them for type = "all".
• predict(type = "metrics") on a resampled model aggregates per case.
• predict() on a CalibratedClassificationRes takes type.
• LightGBM hyperparameters use LightGBM's own names: bagging_fraction, bagging_freq, and boosting replace subsample, subsample_freq, and boosting_type.
• setup_LightGBM(boosting = , data_sample_strategy = ) select DART and GOSS, with their parameters.
• setup_LightGBM() exposes LightGBM's tree regularization, binning, missing-value, categorical, constraint, sampling, quantization, and determinism parameters.
• setup_LightGBM() exposes objective parameters (alpha, tweedie_variance_power, fair_c, poisson_max_delta_step, sigmoid, boost_from_average, reg_sqrt), each applicable only under the objective that uses it.
• setup_LightCART(), setup_LightRF(), and setup_LightRuleFit() expose the LightGBM parameters that apply to each.
• setup_LightRuleFit() names its GLMNET-step parameters alpha_glmnet and lambda_glmnet.
• Printing a Hyperparameters object summarizes unset, tunable, and fixed hyperparameters.
• Property schemas carry x-rtemis.group, naming the declaration group of each hyperparameter.
• conformal() returns prediction intervals or label sets with finite-sample coverage for any supervised model.
• setup_SplitConformal(), setup_CVPlus(), and setup_CQR() configure split conformal, CV+, and conformalized quantile regression.
• Conformal classification sets take score {"APS", "LAC"}.
• conformal_metrics() reports coverage and interval width or set size.
• predict() on a Ranger model leaves the caller's RNG state unchanged.
• Execution configs take shared_memory {"auto", "none", "always"} to pass training data to local workers through 'mori'.
• Execution configs take n_workers_outer, n_workers_tuning, and n_workers_algorithm to assign workers per level.
• setup_MiraiExecution() takes warm_workers to load rtemis in each worker as the pool starts.
• Workers start once per train() call and are reused across dispatches; the execution graph times their startup in a worker_pool node.
• Grid search results are identical across backends and worker counts; tuning results change for algorithms that draw random numbers.
• 'parallelly' moves from Suggests to Imports; 'futurize' and 'future.apply' are no longer dependencies.
• preprocess() modifies a data.table by reference and runs substantially faster.
• setup_Preprocessor(holidays = ) selects the holidays add_holidays flags.
• preprocess() creates date and holiday features before encoding and scaling them.
• setup_Ranger(split_select_weights = ) accepts a single vector applied to every tree.
• setup_Resampler(id_strat = ) takes a column name; train() groups resamples by it and excludes it from the features.
• read() reads Parquet files with Arrow string_view and binary_view columns.

1.3.5

• rtemis is licensed under BSD 3-clause, replacing GPL (>= 3).
• train() parallelizes outer resampling.
• Execution configs take seed, a master seed for computation separate from the resampling seed; an unset seed is drawn and recorded.
• Each outer fold runs on its own RNG substream, so results match across sequential and parallel runs; results for stochastic algorithms under outer resampling change.
• train() leaves the caller's RNG state unchanged when a seed is in effect.
• A parallel run records the same execution graph as a sequential one.
• explain() computes per-case Shapley contributions for every supervised algorithm, exactly where the model structure allows and by KernelSHAP through 'shapr' otherwise.
• setup_SHAP() configures the explanation estimator, perturbation, scale, background, and coalitions.
• explanation_methods() lists, per algorithm, the estimator "auto" selects and whether it is exact.
• shap_case(), shap_long(), shap_by_level(), and get_varimp() read an explanation.
• explain() on a SupervisedRes takes type {"avg", "all"}.
• Shapley contributions are reported per input column, summed over encoded columns.
• reconstruct() returns a decomposition's reconstruction of the data in input units, for PCA, ICA, and NMF.
• decomp() stores reconstruction and component metrics in Decomposition@metrics and the run record.
• decomp_metrics() computes out-of-sample decomposition metrics.
• decomposition_traits() lists each decomposition algorithm's properties; available_decomposition(traits = TRUE) prints them.
• applicable_metrics() lists which metrics apply to which decomposition algorithm.
• Decomposition and Clustering record a data fingerprint, reported in run records.
• NMF scores are each case's non-negative coefficients on the fitted basis; NMF results change.
• setup_NMF(method = ) selects the NMF::nmf() method.
• ICA apply_decomp() reproduces the fitted components under row_norm = TRUE; ICA scores from apply_decomp() change.
• fitted_config() returns a Preprocessor's config with its learned values filled in.
• 'htmltools' moves from Imports to Suggests, needed only by draw_leaflet().
• 'openssl' replaces 'digest' as a dependency.
• setup_MLP() configures a multilayer perceptron through 'torch' for regression and classification.
• setup_MLP() takes explicit hidden_units or generates them from shape, shape_layers, and shape_max_units; hidden_units is tunable, one architecture per candidate.
• MLP encodes categorical features as learned embeddings or one-hot columns, and centers and scales numeric features.
• MLP stops early on dat_validation and keeps the best epoch's weights.
• A fitted MLP can be saved and reloaded.
• MLP uses CUDA when available and the CPU otherwise; device = "mps" is supported on request.
• Printed hyperparameters show search spaces inline, tagged <tune>.

1.3.4

• Tuning requires tune_over(): setup_LightRF(max_depth = tune_over(3L, 4L, 5L)). A bare vector is a value, and max_depth = 3:5 is an error naming the replacement.
• Serialized search spaces are tagged objects, {"candidates": [3, 4, 5]}, with at least two candidates.
• Vector-valued hyperparameters can be tuned.
• apply_preprocessor() codes factors against the training levels for factor2integer and one-hot encoding, so predictions on new data use the training encoding.
• factor2integer codes are integer.
• scale and center skip columns coded by factor2integer.

1.3.3

• setup_SuperLearner() configures the cross-validated stacked ensemble of van der Laan, Polley & Hubbard for regression and binary classification.
• setup_ModalityStacking() stacks base learners, each bound to one group of features.
• setup_ConditionalSuperLearner() routes each case to one expert from a library through a learned oracle (Valdes, Interian, Gennatas & van der Laan, 2022).
• setup_NNLS() configures non-negative least squares through 'nnls', the default stacking meta learner.
• setup_MARS() configures Multivariate Adaptive Regression Splines through 'earth' for regression and classification.
• train(weights = ) accepts a numeric vector.
• A SuperConfig weights column supplies case weights and is excluded from the features.
• read_config() validates configs without calling the rtemis CLI; options(rtemis.validate = ) and options(rtemis.cli = ) are removed.

1.3.2

• Isotonic calibration keeps probabilities at least 1 / (2 * n) from 0 and 1.
• calibrate(hyperparameters = NULL) selects setup_Isotonic().
• Grid search leaves unset any hyperparameter that does not apply to the rest of a combination, and collapses duplicate combinations.
• tuning_grid() returns the combinations a grid search would fit.
• Published schemas declare hyperparameter dependencies in x-rtemis.applies_when.
• setup_SPLS() configures Sparse Partial Least Squares through 'spls'.
• setup_KNN() configures weighted k-Nearest Neighbors through 'kknn'.
• setup_BART() configures Bayesian Additive Regression Trees through 'stochtree'; se() returns the posterior standard deviation of the mean function.
• setup_HAL() configures the Highly Adaptive Lasso through 'hal9001'.
• train() projects a HAL basis size before fitting, warns past one million, and aborts past max_basis.
• setup_MonotonicHAL() configures a monotonic non-decreasing Highly Adaptive Lasso, usable as a calibrator.
• available_calibration() lists the algorithms usable as calibration maps.
• CalibratedClassification@calibrator and CalibratedClassificationRes@calibrator record the calibrator used.
• plot_varimp(measure = ) selects among a model's importance measures.

1.3.1

• Predicted probabilities are a matrix with one row per case and one column per class; binary outcomes carry a single column for the positive class.
• train(), calibrate(), setup_SuperConfig(), and setup_SuperConfigLive() no longer take algorithm; the setup_*() function names the algorithm.
• se(mod, newdata) computes standard errors on demand, replacing the @se_* properties of Regression and RegressionRes.
• The .list_to_*() reconstructors reject undeclared keys and suggest the nearest valid property.
• decomp(x, config) uses config@features.
• A classification result's positive_class is NULL when the outcome is not binary.
• train(), decomp(), and cluster() write a <prefix>.record.json run record to outdir, stating every resolved value with its origin, provenance, and a data fingerprint.
• record() returns a run record; write_record() writes one.
• Run records list each fitted model under folds, with its tuning results, and include headline metrics per sample.
• Supervised, SupervisedRes, Decomposition, and Clustering store the config they were given; read_config() rejects a run record.
• Fitted models report hyperparameter values resolved during training, such as LightGBM's nrounds.
• Published schemas include a record.json beside each config schema.
• RegressionMetrics, ClassificationMetrics, and their resampled counterparts have typed, validated tables and published schemas.
• Per-class metrics name their outcome level in a level column.
• Classification metrics carry confusion_long, a long-format confusion table.
• to_json() emits the properties the published schemas declare.
• Factor-valued properties serialize as {levels, codes}.
• write_config() and read_config() support preprocessor configs.
• JSONSchema_to_S7() builds an S7 class from a schema generated by S7_to_JSONSchema().
• Hyperparameters bounded by the training data are checked before any model is fit.
• setup_LightRF(feature_fraction = NULL) derives the value from the data.

1.3.0

• Configuration classes declare typed, validated properties from which their JSON Schemas are generated.
• Optional config properties use NULL as the only unset value and reject zero-length values.
• S7_to_JSONSchema(), S7_dispatcher_JSONSchema(), and write_JSONSchema() generate and write JSON Schemas from S7 classes.
• setup_Preprocessor(impute_type = ) lists its choices in the signature.

1.2.8

• session_timeline() flattens a SupervisedSession execution graph into a timeline table; session_kind_colors() provides a shared color map.
• Progress reporting uses rtemis.core::progress_lapply(), including for parallel tuning; 'cli' is no longer a dependency.
• train(dat) defaults to Ranger.
• read() removes duplicates through preprocess(); its make_unique argument is renamed remove_duplicates.
• apply_preprocessor() applies a trained Preprocessor to new data, replacing preprocess(x, Preprocessor).
• preprocess() accepts setup_Preprocessor() arguments directly, e.g. preprocess(x, scale = TRUE).

1.2.7

• Added DecomposeConfig and ClusterConfig pipeline-recipe classes with setup_DecomposeConfig() / setup_ClusterConfig(), mirroring SuperConfig: they bundle a data path, the algorithm config (DecompositionConfig / ClusteringConfig), and an output directory.
• decomp() now accepts DecomposeConfig objects.
• cluster() now accepts ClusterConfig objects.
• Added outdir arg to decomp() and cluster().
• Added read_config() & write_config() with support for the new supervised, decompose, cluster schemas.

1.2.6

• SupervisedRes now records preprocessor_config and decomposition_config; updated repr.
• Added early input validation: column-type check in check_supervised, new check_numeric_or_factor(), and a check that the requested decomposition exists.
• decomp() now reports the number of features and components.
• Exported show_color_key().
• repr moved to rtemis.core; metric acronyms now capitalized in console output.
• Extracted roc_curve().

1.2.5

• Adopted the rtemis.core condition system (rtemis.core::abort() / warn()); documentation now links to rtemis conditions.
• read() now errors if the file does not exist.
• Improved sanitize_path().
• Initial SupervisedSession support.

1.2.4

• Added nanoparquet support for reading and writing data (added to Suggests).
• Added default_n_workers(), used in .onAttach().
• Added numeric_features() generic and methods.
• Added features property to DecompositionConfig for use within train().
• WASM-safe parallel-worker detection.
• Moved shared utilities to rtemis.core.

1.2.3

• Added decomposition_config support to SuperConfig, SuperConfigLive, and train(), with new apply_decomp methods where supported.
• Converted Regression and Classification metric field names to lower case.
• Added description field to to_json() output.
• Added verbosity argument to the describe() S7 generic and methods.

1.2.2

• Added set_positive_class(). Can be used directly by users. Used by rtemis.server and rtemislive to pass the positive case from the UI to rtemis.
• Added positive_class field to SuperConfigLive.
• Added aggregated confusion matrix to resampled classification results (ClassificationMetricsRes); moved Confusion_Matrix out of the metrics object.
• Added progress argument to train() to allow a callback for rtemis.server.
• Added get_varimp() method.
• Added rtemis.core to Imports.

1.2.1

• Added the package name to S7 class definitions.
• Exported additional internals required by rtemis.server.

1.2.0

• Add rtemis.server support:
• New SuperConfigLive S7 class for server-based training configuration.
• New set_msg_sink(), get_msg_sink(), with_msg_sink() functions to capture and redirect rtemis console messages.
• New to_json() S7 generic to convert rtemis objects to JSON-serializable lists.
• Add verbosity argument to predict_super(); remove ....
• Add names() S7 method for Theme objects.

1.0.1

• Introduce VariableImportance S7 class to represent variable importance data, allowing for more than one measure of importance per model and update all relevant classes and methods.
• Calculate Partial_Effect_Variance as variable importance measure for GAM models
• Add execution_config argument to internal train_ method and use it in LightRuleFit to propagate to LightGBM and GLMNET calls.

1.0.0 First CRAN release

rtemis documentation built on Oct. 20, 2026, 1:06 a.m.